[laudney]

Part of #1927.

Some checkpoints never emit a reasoning block, but their chat templates still read enable_thinking. The OCR models in #1927 are like this. Today, the server splits their replies as if they could think: with thinking on, which is the default, the whole OCR text lands in reasoning_content, and content stays empty. This PR adds one engine flag for such models. A family that sets it gets the template's own prompt rendering, and its reply is always answer content.

	Before (main)	After (this PR)
Reply of a no-reasoning model with thinking on (the default)	all text in reasoning_content, content empty	all text in content
Prompt rendering under enable_thinking	follows the template	unchanged: follows the template
Harmony channel markers in such a reply	parsed into channels	kept as answer text, streamed and whole
Existing families (Info.reasoning defaults to true)	—	unchanged on both reply paths
How a native HIP family opts out	—	pub const reasoning = false;
Places that combine thinking with this capability	—	one, chat.prepare (Prepared.reasoning)

Changes

• engine_api.Info.reasoning: bool = true.
• chat.prepare keeps thinking for prompt rendering. It computes reasoning = thinking and srv.info.reasoning once and stores it on Prepared. Each reply-side use reads it:
  • feature checks
  • the think budget and reserve
  • logprobs admission
  • the think-end marker
  • the reasoning split and its streamed counterpart
  • the length warning for an empty reply
• With reasoning = false, the Harmony fallbacks (parseHarmony in finish and streamingVisible in the stream path) do not run.
• native/hip.zig: a family's pub const reasoning = false; sets the info flag. A family without that declaration keeps true.

Gates

• hf_eos_http_test.zig, through the real HfText and the OpenAI handler, both streamed and whole:
  • reasoning = false under default, explicit-on and explicit-off thinking: the prompt follows the template, the reply is all content, and reasoning_tokens is 0.
  • The default true under the same three settings: the split works exactly as before.
  • A Harmony reply (<|channel|>analysis<|message|>plan<|channel|>final<|message|>OK) with reasoning = false stays complete answer content. This test fails if the stream-path guard is removed.
• native/hip.zig: the registry test shows that a declared family gets false and every other family keeps true.
• zig build test: 465/465 steps. zig build test -Dhip=true: 492/492 steps. test-cuda-host and the native-hip build pass.
• Lean checker: zero new problems over the 425 baseline. chat.zig stays at 600 lines.

[ashhart]

Thanks @laudney, reviewed: this lands. Info.reasoning defaults to true, so every path on main stays exactly as it was (reasoning equals thinking, and the Harmony parse and the think-close id are unchanged). No registered family sets it false yet, as your test checks, so it only takes effect once the #1927 OCR families declare it. Its red zig and zig-windows checks aren't from this PR: main has failed qwen4_cuda_attention_host.cpp (fn_attn_merge_body) since #1739, and #1935 fixes that. Once #1935 is in, we'll rebase this onto main, let its CI run, and land it. You don't need to push anything.

[ashhart]

Thank you @laudney. This is merged into main with your commits under your own name, so you're listed as a TensorFold contributor. It ships in the next release, credited to you in the changelog and the release notes.

[ashhart]

Thanks @laudney! This is in TensorFold 1.2.2, out now: https://github.com/ashhart/TensorFold/releases/tag/v1.2.2